856123
research-article2019
ARPXXX10.1177/0275074019856123The American Review of Public AdministrationBullock
Article
Artificial Intelligence, Discretion, and
Bureaucracy
American Review of Public Administration
2019, Vol. 49(7) 751­–761
© The Author(s) 2019
Article reuse guidelines:
sagepub.com/journals-permissions
https://doi.org/10.1177/0275074019856123
DOI: 10.1177/0275074019856123
journals.sagepub.com/home/arp
Justin B. Bullock1
Abstract
This essay highlights the increasing use of artificial intelligence (AI) in governance and society and explores the relationship
between AI, discretion, and bureaucracy. AI is an advanced information communication technology tool (ICT) that changes
both the nature of human discretion within a bureaucracy and the structure of bureaucracies. To better understand this
relationship, AI, discretion, and bureaucracy are explored in some detail. It is argued that discretion and decision-making are
strongly influenced by intelligence, and that improvements in intelligence, such as those that can be found within the field of
AI, can help improve the overall quality of administration. Furthermore, the characteristics, strengths, and weaknesses of
both human discretion and AI are explored. Once these characteristics are laid out, a further exploration of the role AI may
play in bureaucracies and bureaucratic structure is presented, followed by a specific focus on systems-level bureaucracies.
In addition, it is argued that task distribution and task characteristics play a large role, along with the organizational and legal
context, in whether a task favors human discretion or the use of AI. Complexity and uncertainty are presented as the major
defining characteristics for categorizing tasks. Finally, a discussion is provided about the important cautions and concerns of
utilizing AI in governance, in particular, with respect to existential risk and administrative evil.
Keywords
artificial intelligence, bureaucracy, discretion, administration, governance
Introduction
In 1980, Michael Lipsky codified the role of the individual
decision maker, the “street-level” bureaucrat (Lipsky, 1980).
Such bureaucrats, including teachers, doctors, case workers,
police officers, fire fighters, garbage collectors, and information service workers, provided government services, benefits,
and punishments directly to the public (often, quite literally,
out on the street). As Lipsky noted, these street-level bureaucrats used their discretion—their decision-making latitude in
implementing government policies in complex and uncertain
problem spaces—to serve the public and solve problems.
Lipsky’s work inspired much scholarly curiosity in the
field of public administration (PA), into the motivations of
these street-level bureaucrats (and their managers) who were
implementing policy. Because bureaucrats exercise discretion when completing their job tasks, it would be helpful to
understand how these decision-making processes work.
Scholars examined differences between government bureaucrats and private sector employees across motivations and
attitudes such as public service motivation (Perry, 1996;
Perry & Wise, 1990), extrinsic motivation (Bullock, Hansen,
& Houston, 2018), and organizational commitment (Bullock,
Stritch, & Rainey, 2015). In addition to this work, PA gave
rise to the field of public management, where the impacts of
managerial decisions on performance were highlighted
(Bullock, Greer, & O’Toole, 2019; Meier & O’Toole, 2001,
2002; O’Toole & Meier, 1999, 2003). This research empirically established the relationship between those working
within government and the performance of government
(O’Toole & Meier, 2011).
In the midst of this wave of scholarly attention devoted
to understanding the motivations, attitudes, and decisionmaking processes of street-level bureaucrats, Bovens and
Zouridis (2002) observed a trend in large, executive, public
agencies, from street-level bureaucracies, to screen-level
bureaucracies, on to systems-level bureaucracies. No longer were bureaucrats simply “out on the street” providing
services to the general public; now much of this work was
being done behind a computer screen. Many decisions were
becoming routinized and guided by computer software and
databases (screen-level bureaucracy), which resulted in
changes in the operation and scope of discretion itself.
Decision-making tasks were guided by software for case
management, performance tracking, and increases in the
1
The Bush School of Government & Public Service, Texas A&M
University, College Station, USA
Corresponding Author:
Justin B. Bullock, The Bush School of Government & Public Service, Texas
A&M University, 4220, TAMU College Station, TX, 77843-4220, USA.
Email: jbull14@tamu.edu
752
oversight and accountability of bureaucrats’ discretion
(Bovens & Zouridis, 2002).
Bovens and Zouridis (2002) maintain that the transition
from street-level bureaucracy to screen-level bureaucracy
was only the first for some bureaucracies. Screen-level
bureaucracies began to give way to even more expansive
uses of information communication technology (ICT) tools,
such as computers, email, and other uses of the Internet and
digital world. Meanwhile, the ICT tools themselves increased
in their capacity to both augment and automate decisionmaking tasks (Kurzweil, 2005, 2013). In some instances,
these increases began to give way to a new form of bureaucracy, the systems-level bureaucracy, characterized by the
decisive role that ICT tools play in the bureaucratic process
for some agencies. No longer were bureaucrats out on the
ground or sitting behind screens inputting data; instead, ICT
tools themselves augmented and automated more and more
tasks, in some cases even replacing the role of expert judgment by human bureaucrats (Bovens & Zouridis, 2002).
In the 17 years since Bovens and Zouridis observed a burgeoning transition to systems-level bureaucracies, the rise in
power of ICT tools, and in particular artificial intelligence
(AI), brings to the forefront classic PA questions about discretion, bureaucracy, and administration. This essay seeks to
explore the impacts of AI on discretion and the potential consequences for bureaucracy and governance.
AI’s use as a tool of governance is growing and this trend is
being discussed by the U.S. Federal Government and debated
by scholars. For example, on June 29, 2018, SpaceX launched
an artificially intelligent robot named CIMON to the
International Space Station to aid NASA astronauts
(Davenport, 2018). In addition, the U.S. General Services
Administration has partnered with a number of U.S. federal
agencies to explore the use of artificially intelligent personal
assistants. The use of AI as part of predictive policing and
facial recognition for surveillance is growing. In “The Rise of
Big Data Policing,” Ferguson (2017) documents the growth of
AI and digital surveillance in law enforcement and predictive
policing. Furthermore, Virginia Eubanks (2018) argues that
the automation of tasks once the purview of street-level
bureaucrats has led to an automation of inequality, giving rise
to 21st-century digital poorhouses. Eubanks highlights how
states’ use of algorithms to make decisions about government
benefits and punishments disproportionately harms the poor.
Recently, the U.S. federal government has made AI a policy priority, issuing reports affecting both federal agencies
and the private market. On December 20, 2016, the Executive
Office of The President released “Artificial Intelligence,
Automation, and the Economy,” signed by the Chair of the
Council of Economic Advisers, the Directors of the Office of
Science and Technology Policy, the Domestic Policy
Council, and the National Economic Council, and the U.S.
Chief Technology Officer (Executive Office of the President,
2016). This report focuses on the emergence of AI in the
economy, but also can be applied to the types of tasks
American Review of Public Administration 49(7)
traditionally in the purview of PA and street-level bureaucrats. “Accelerating AI capabilities,” it states,
will enable automation of some tasks that have long required
human labor. Rather than relying on closely-tailored rules
explicitly crafted by programmers, modern AI programs can
learn from patterns in whatever data they encounter and develop
their own rules for how to interpret new information. This means
that AI can solve problems and learn with very little human
input . . . . This (the combination of AI & robotics) will permit
automation of many tasks now performed by human workers
and could change the shape of the labor market and human
activity. (Executive Office of the President, 2016, p. 8)
Even more recently, Executive Order (EO) 13859,
“Maintaining American Leadership in Artificial Intelligence,”
was issued on February 14, 2019. This EO notes that AI will
affect essentially all federal agencies, and that those agencies
should pursue six strategic objectives in response, including
promoting sustained investment in AI Research &
Development, enhancing access to high-quality federal data,
reducing barriers to the use of AI technologies, ensuring
technical standards minimize vulnerabilities to attacks from
malicious actors, training the next generation of AI researchers, and developing a strategic plan to protect the United
States’ advantage in AI and technology critical to U.S. interests (Executive Office of the President, 2019). The U.S. federal government has begun to prioritize issues of AI and
improving governance. PA scholars should do the same.
This article seeks to provide an exploration of the role AI is
playing in changing the nature of bureaucratic discretion. To
better understand this changing nature of discretion, discretion
itself needs to be better situated in a broader discussion of PA
and intelligence. Thus, an argument is presented that decisionmaking, as Herbert Simon (1947) argued, is at the heart of
administration, and that intelligence is the basic capacity for
decision-making. In this context, human discretion is characterized as human bureaucrats engaging in decision-making (in
which uncertainty is present) for a complex task about how
best to deliver government services, benefits, and punishments. Once this argument is elucidated, then the human terrain for decision-making and quality of discretion are
elaborated upon. Then, some of the basic characteristics of AI
are presented. From here, arguments are made about what the
rise of AI means for the future of discretion within bureaucracy broadly and more specifically about the impact AI has
on discretion within systems-level bureaucracies. Finally, the
article concludes with some concerns and caution about how
AI should be used intelligently as a tool of governance.
Administration, Discretion, and
Intelligence
Herbert Simon noted that decision-making is at the very heart of
administration (Simon, 1947). Street-level public servants use
their decision-making latitude—their discretion—to conduct
Bullock
the basic work of administration. Once legislation is passed,
bureaucratic discretion occurs at all levels, from the details of
policy implementation, to agenda- and priority-setting, to the
design of public agencies and organizations, and to the actual
delivery of services on the ground (Busch & Henriksen, 2018).
Quality of administration can be characterized, in part, by
how effectively public administrators use their discretion to
achieve policy goals. An understanding of the individual
decision-making process is an important feature of the study
of administration. But what drives decision-making capacity? Can some construct be used as a guide to the quality of
decision-making, and thus the quality of discretion? While
both motivation and goal alignment of the individual bureaucrat play large and important roles, psychologists, neuroscientists, and AI scholars argue that intelligence is the construct
that helps to define the overall capacity for the quality of
decision-making (Dreary, Penke, & Johnson, 2010; Piaget,
2005; Russell & Norvig, 2016).
This essay adopts Max Tegmark’s (2017) definition of
intelligence, “the ability to accomplish complex goals.”
Tegmark argues that this definition is broad enough to
encompass many of the competing definitions of intelligence, which include “capacity for logic, understanding,
planning, emotional knowledge, self-awareness, creativity,
problem solving and learning” (loc 919, Tegmark, 2017).
This definition also allows for many types of intelligence, in
pursuit of many different types of goals and gives a common
construct that can be assessed across agents that pursue
these goals (with a bonus characterization of substrate independence, which is helpful in thinking about intelligence
across organic and nonorganic entities; Bostrom, 2014;
Tegmark, 2017).1
Administration is the attempt to achieve complex organizational goals through a decision-making process; intelligence is the ability to accomplish complex goals. Thus, those
interested in the study of administration should be concerned
with advances in intelligence, and these advances abound,
particularly in the realm of AI. However, before discussing
AI, it is useful to consider the strengths and limitations inherent in human intelligence, and hence human discretion.
Human Discretion in Administration
Human discretion exists at the heart of administration.
Humans created and maintained the administrative state, and
they remain its designers, leaders, and implementers. Given
this history of human discretion in administration, what is
known about the decision-making process and how well
humans perform as actors with discretion in the delivery of
public services? To treat this topic fully, in all its nuances,
would require a book. In fact, numerous books have been
written, for example, on how street-level bureaucrats use
their discretion (Lipsky, 1980); entire textbooks explain how
humans design, manage, and run public organizations
(Rainey, 2014). A recent wave of research in behavioral PA
753
looks to incorporate what behavioral economists and psychologists have discovered about the systematic biases and
limitations to human rationality, and how this plays out in the
field.2
A brief treatise must suffice. Ample empirical evidence
suggests that human public administrators are quite capable
of efficiently, effectively, and equitably completing a wide
range of administrative tasks (Rainey, 2014). Early work in
PA examined a core set of key administrative tasks known as
POSDCORB: planning, organizing, staffing, directing, coordination, reporting, and budgeting (Gulick, 1937), and these
tasks have been accomplished by humans throughout history.
Administrative tasks rely on the ability to flexibly coordinate
across a number of domains; such tasks are quite suitable to
human strengths (Harari, 2016). Administrators engage in
networking, politicking, and innovating (Borins, 1998, 2008;
Dunn, 1997; Goldsmith & Eggers, 2004; Isett, Mergel,
LeRoux, Mischen, & Rethemeyer, 2011; Linden, 1990), all
forms of discretion. They are often motivated to serve the
public, find meaning in helping others, and work to help efficiently and effectively make use of public resources to provide public goods (Bullock et al., 2018; Bullock et al., 2015;
Perry & Wise, 1990). They cleverly satisfice, coordinate, and
compromise to handle the multitude of tasks required of
them (Simon, 1947).
But the human discretion picture is not always a rosy one.
Humans are not rational automatons (Kahneman, 2011).
They are riddled with a set of systematic cognitive biases and
subjective probability weights (Bazerman, 2002; Kahneman
& Tversky, 1972; Wilkinson, 2008) and succumb to planning
fallacies and to loss aversion (Kahneman & Tversky, 1979).
They are susceptible to corruption (Klitgaard, 1988). They
become demotivated and do not perform their job tasks well.
They award improper and fraudulent benefits and deny benefits to the deserving (Greer & Bullock, 2018; Wenger &
Wilkins, 2008). They use their power and influence to cause
harm to the weak. They are often overconfident in their own
abilities and quickly run into limits in their capacity to process large amounts of data (Kahneman, 2011). While humans
have certainly improved their administrative abilities over
time to match the changing demands of modern society, there
is no question that human administrators are limited in their
capacities and deeply flawed in their pursuit of efficiency,
effectiveness, and equitable administration of the law.
Over time, ICT tools have aided humans in refining and
improving their discretion, and thus the overall quality of
administration. Early ICT included basic systems of written
accounting, but these tools have progressed to complex technological systems in which AI systems do more and more of
the tasks that were once in the domain of human bureaucrats.
This march of improvement in the capacity of ICT tools allows
for augmentation and automation of tasks and continues to
encroach into the space of work once completed solely by
human administrators (Busch & Henriksen, 2018). In addition, ICT progress presents its own challenges (Barth &
754
Arnold, 1999; Bovens & Zouridis, 2002; Simon, 1997), which
will be further elaborated upon later in this essay. While these
challenges are not necessarily new (automation and advances
in computing have made numerous administrative tasks obsolete over time), advances in AI appear especially dramatic.
What began as the automation of physical efforts and relatively straightforward data entry is evolving to the domains of
cognitive and analytical tasks (Brundage et al., 2018).
AI: Basic Characteristics and
Opportunities for Administration
AI can be thought of as the ability of a nonorganic, mechanical entity to accomplish complex goals (Tegmark, 2017). As
mentioned earlier, this implies that intelligence is substrate
independent: It can exist as a biological process informing
human actions or as a mechanical process guiding an automaton. AI literature distinguishes between “weak” or “narrow
AI” and “strong” or “general AI.” Weak or narrow AI can
learn to complete a small set of tasks or one particularly challenging goal. Strong or general AI can learn across a broad
range of challenging complex tasks across almost any
domain (Nilsson, 2009). There is much discussion about the
rising abilities of deep neural networks and generative adversarial networks (Bostrom, 2014), but for the purposes of this
article, this will suffice: Narrow AIs are being continuously
improved and are increasingly able to accomplish tasks that
were once most effectively, efficiently, and equitably completed by humans.
As AI progresses, potential applications to the work of
administration and governance will multiply, though governments often do lag in their adoption of new technologies. In a
resource-constrained environment, public organizations dare
not invest in large updates to their computing hardware and
software. However, as these costs fall, and AI gains in replicability, scalability, and efficiency (Brundage et al., 2018),
the same pressures that delay government technological
adoption may make these tools attractive investments in organizational efficiency. It seems that with current technologies,
almost two decades after Bovens and Zouridis’ observation of
the beginnings of systems-level bureaucracies, many tasks
carried out by human actors may already be more efficiently
and effectively carried out by AI. This suggests that tasks
completed by human street-level bureaucrats, across many
domains, may be augmented or automated by AI.
AI may bring increasing levels of effectiveness and efficiency, but threats to governance remain, such as impacts on
equity, accountability and democracy (Barth & Arnold,
1999), along with concerns of managerial capacity and political legitimacy (Salamon, 2002). Given competing governance goals and values, these opportunities and threats
present important questions for how discretion should be
employed within the context of public organizations. These
questions permeate all organizational levels of governance
tasks and are directly addressed in the following sections
(Busch & Henriksen, 2018).
American Review of Public Administration 49(7)
The rise of AI’s capabilities will result in the augmentation
and automation of cognitive tasks. Market analysts estimate
that advances in AI and robotics are likely to lead to further
automation in at least 10 different task areas including: personalized products, targeted marketing, fraud detection, supply chain and logistics, credit analysis, construction, food
preparation and service, autonomous vehicles, automated
checkout, and health care provision (Helfstein & Manley,
2018). Consider a few examples of these task areas and their
relationship to administration and governance. A wave of disruption is already cresting, and unless the field addresses these
challenges preemptively, it is easy to imagine a violent crash.
Personalized products could dramatically change how
taxes are collected through automation of data collection
from employer and state databases. These same products
could help identify those who are legally entitled to sets of
social services. Targeted marketing could certainly be used
to help ensure that citizens receive all relevant public service
announcements. Fraud detection could be improved dramatically, particularly in entitlement and social insurance programs, eliminating street-level bureaucrat positions in fraud
management. Supply chain and logistics could be streamlined across a host of government business practices, thus
potentially improving governance efficiency. Credit analysis
could be improved for governments of all sizes, with the creation of more effective tools to accurately price credit and
determine creditworthiness.
Automated food preparation and service could modernize
cafeterias for public organizations throughout the country.
Autonomous vehicles will likely replace bus drivers, or at
least change the nature of their jobs, and revolutionize transportation across government agencies. The public information officers mentioned by Bovens and Zouridis may become
obsolete, as the design of automated checkout systems
improves and the public’s comfort with them increases. The
health care provider industry (another category of nonroutine, nonlegal, and complex work mentioned by Bovens and
Zouridis) seems to be in the midst of its own AI revolution;
this could have dramatic implications for the delivery of public health services (Topol, 2019). These are just some examples of the breadth of PA tasks that could be automated or
augmented by AI. The expansiveness of this task terrain
presents a question for scholars of PA: What is the future of
discretion for street-level bureaucrats?
The Future of Street-Level
Bureaucrats’ Discretion in Bureaucracy
Bovens and Zouridis (2002) noted the changing character of
large, executive agencies, stating that many of these “public
agencies of the welfare state appear to be quietly undergoing
a fundamental change of character internally. Information
communication and technology (ICT) is one of the driving
forces behind this transformation” (p. 175). They note that a
number of large executive organizations made a shift from
traditional street-level bureaucracy to screen-level and, for at
Bullock
least two agencies in the Netherlands, there was a shift
toward systems-level bureaucracies.
Bovens and Zouridis (2002) characterize systems-level
bureaucracies as those in which
ICT has come to play a decisive role in the organizations’
operations. It is not only used to register and store data, as in the
early days of automation, but also to execute and control the
whole production process. Routine cases are handled without
human interference. Expert systems have replaced professional
workers. Apart from the occasional public information officer
and the help desk staff, there are no other street-level bureaucrats
as Lipsky defines them. (p. 180)
There are no other street-level bureaucrats.
While Bozens and Zouridis are quite cautious in their suggestion of the expanse of systems-level bureaucracies (“It
remains to be seen whether similar transformations can be
observed in non-legal, non-routine, street-level interactions
such as teaching, nursing, and policing” (p. 180), they argue
that bureaucracies that do transition to systems-level will
only employ three groups of people: (a) those active in the
data processing process, (b) managers of the production process, (c) and those that help clients interface with the information system. This shift to systems-level bureaucracies
may greatly reduce the scope of administrative work that
would be completed by humans, leaving very little, if any,
work left for the traditional street-level bureaucrat.
Building from Bovens and Zouridis’ argument, and the
work of other scholars, there is a growing debate in PA about
whether ICT has a curtailing or enabling influence on the
discretion of human street-level bureaucrats. Buffat (2015)
highlights this debate:
On the one hand, initial research considered that street-level
discretion decreases or disappears in the case of large bureaucratic
informatisation. Since an insistence is put on the negative impact of
ICT over discretion, the label “curtailment thesis” is relevant. On
the other hand, other studies point out more nuanced effects of ICT.
These studies indicate that new technologies constitute only one
factor among others shaping street- level discretion and that they
provide both frontline agents and citizens with action resources.
This why we label this orientation the “enablement thesis.” (p. 151)
Buffat (2015) notes that Snellen (1998, 2002), Zuurmond
(1998), and Bovens and Zouridis (2002) make versions of the
curtailment argument. Buffat summarizes these curtailment
arguments by stating that the scholars believe that “ICT has a
negative and curtailing effect on frontline discretion. In computerized public service delivery, street-level bureaucrats partially or totally lose their discretionary power” (p 153). Buffat
(2015) notes four weaknesses to the curtailment argument: (a)
it implies technological determinism, (b) discretion is too narrowly defined, (c) empirical limitations, and (d) an insufficient interest in the concrete uses of technologies by frontline
workers. He offers a different hypothesis, which is labeled the
“enablement thesis.”
755
Buffat notes evidence from several studies including a
subsidy allocation process in the Netherlands (Jorna &
Wagenaar, 2007), ICT and policy discretion in a Swiss unemployment fund (Buffat, 2011), implementation of French
agricultural policy (Weller, 2006), and the use of a French
social aid website (Vitalis & Duhaut, 2004). These studies
lead Buffat to claim
discretion is not suppressed at the frontline despite ICT tools and
continues to exist in the daily street-level activity. This result is
linked to contextual factors such as the inability of ICT tools to
capture the whole picture of frontline work and choices, limited
resources for managers to control time and attention, and work
organization or the skills possessed by street-level agents.
Analytically speaking, this means that technology (and its use)
is only one of the factors influencing the discretion of frontline
agents and that a variety of non-technological factors shape it as
well. This is why no unilateral effects of technology can be
assumed. (Buffat, 2015, pp. 156-157)
Busch and Henriksen (2018) come to a somewhat different
conclusion. They use Buffat’s constructs of enablement and
curtailment and find evidence for both in their literature
review on digital discretion. Building from Buffat’s call for a
better understanding of the contextual factors contributing to
the diffusion of digital discretion, and informed by their own
literature review, Busch and Henriksen identify 10 contextual factors nestled within four levels of analysis. These factors include: at the macro-level policy maker: formulation of
rules; at the meso-level street-level bureaucracy: formulation
of organizational goals, formalization of routines, and interagency dependency; at the micro-level street-level bureaucracy: professionalization, computer literacy, decision
consequences, information richness, and relational negotiations; and at the technology level: features. Busch and
Henriksen also identify and discuss six technologies from
empirical studies, including telephone, multifunctional computer, database, website, case management system, and an
automated system.
From their literature review, Busch and Henriksen question
Lipsky’s claim that “the nature of service provision calls for
human judgement that cannot be programmed and for which
machines cannot substitute” (Lipsky, 1980, p. 161). Their
response addresses two questions: (a) can digital discretion
(defined by the authors as “increased use of technologies that
can assist street-level bureaucrats in handling cases by providing easy access to information about clients through channels
such as online forms, and automating parts of or whole work
processes”) cause a value shift in street-level bureaucracy and
(b) under what conditions can this shift occur? They find that
there is an increase in diffusion of digital discretion in governance. They argue that the empirical evidence to date (which
they note is sparse) suggests digital discretion “seems to be
more suitable for strengthening ethical and democratic values,
and in general fails in its attempts to strengthen professional
and relational values” (Busch & Henriksen, 2018, p. 18).
The authors argue that digital discretion is diffusing, it is
756
influencing street-level bureaucrats, and ICT is sometimes
substituting for human judgment. Finally, the authors note that
the overall scope of street-level bureaucracy is decreasing.
The authors conclude that
while certain types of street-level work seem to avoid extensive
changes due to ICT, it makes more and more sense to talk about
digital bureaucracy and digital discretion, since an increasing
number of street-level bureaucracies are characterized by digital
bureaucrats operating computers instead of interacting face-toface with their clients. (Busch & Henriksen, 2018)
However, they do note that one technology in particular has
the potential to automate even nonroutine street-level work
such as nursing, teaching, and policing. They state that
artificial intelligence is a technology that has developed rapidly
in recent decades describing computers that can act as
autonomous agents and approximate the human brain. For
example, IBM’s Watson can now better identify symptoms of
diseases than experienced physicians, and AI has proved to
conduct better assessments of English essays than teachers.
(Busch & Henriksen, 2018)
What do these arguments suggest for the future of streetlevel bureaucrats’ discretion within bureaucracy? The first
observation is that AI enables human discretion across some
tasks and curtails discretion across others. In this way, as
Buffat (2015) and Busch and Henriksen (2018) highlight, the
context of the task greatly matters. Busch and Henriksen provide a number of contextual organizational factors. The second observation is that not only does the context matter, but
the type of task itself matters. Here, Perrow (1967) provides
some useful insight by classifying work tasks along two
important dimensions: the frequency with which the task
requires deviations from normal procedures and the degree
to which the task is analyzable or can be broken down into a
rational systematic search process. While the organizational
contextual factors provide guidance of where to look for discretionary tasks within bureaucracy, this task classification
can be very helpful if two assumptions are added: (a) tasks
that rarely deviate from normal procedures are less complex
and more routine, and are therefore more likely to be automated or augmented by AI, compared with those that often
deviate from normal procedures and are less routine, and (b)
the more analyzable a task is the less uncertainty there is
around that task and the more likely it is to be automated or
augmented by AI, and the less analyzable a task, the less
likely it is to be automated or augmented by AI. The greater
the deviations from normal procedures, the more complex
the task, which makes it more difficult to automate. The less
analyzable a task, the greater the uncertainty, which also
makes a task difficult to automate. Thus, tasks that are high
in complexity (more deviations from the norm) and high in
uncertainty (less analyzable) should most likely remain as a
discretionary task to be completed by humans. Furthermore,
American Review of Public Administration 49(7)
tasks that are lower in complexity (fewer deviations from the
norm) and lower in uncertainty (more analyzable) are more
likely to be completed by a machine.
While human discretion remains most important in tasks
where complexity and uncertainty are high, given humans’
weaknesses in situations of uncertainty, AI also maintains a
comparative advantage to humans with dealing with uncertainty. Thus, AI likely has a relative advantage for tasks that
are high in uncertainty but low in complexity. Conversely,
when a task is low in uncertainty but high in complexity,
humans can identify highly complex and abstract patterns
across task sets (a more general intelligence that still escapes
AI), so humans likely retain a relative advantage in this
domain. Figure 1 illustrates this relationship in a 2 × 2
matrix. It is important to consider the task itself in the diffusion of digital discretion and AI. Organizational context matters as well, but the complexity and uncertainty around any
given task will also influence whether the task likely remains
in the domain of human discretion.
A final observation is that the complexity and uncertainty
of a task are likely to vary along the contextual factors highlighted by Busch and Henriksen and across different types of
governance organizations and organizational units within a
governance entity. For example, imagine the structure of a
local government of a mid-size city. This city may have a
transportation department, a police department, a fire department, an accounting department, a legal department, a human
resources department, a library services department, and a
parks and recreation department, among many others. While
some tasks and task distributions will be similar across these
departments, others will vary significantly based upon the
work of the particular unit. For example, fire fighters and
police officers encounter complex and uncertain task situations in their interactions with the citizenry; such complex
and uncertain encounters may result in extreme deviations
from normal procedures and limits in the ability to fully analyze the situation. By contrast, the task distribution for work
tasks in an accounting office, a human resources department,
and a library services department may consist of more tasks
with fewer deviations from normal procedures (less complex) and that are more analyzable (less uncertain).
It is clear that the scope of street-level bureaucrats’ discretion is decreasing in some domains, and the overall role
human discretion plays in administration is changing.
Discretion is curtailed and automated across some tasks and
enabled and augmented across others. This is evident in the
contextual analysis provided by Busch and Henriksen (2018)
and in the argument provided here on task complexity and
uncertainty. In addition, the distribution of types of tasks
completed by a governance organization contributes to the
degree to which the organization is likely to shift from being
a street-level bureaucracy to a screen-level bureaucracy and
on to a systems-level bureaucracy. Systems-level bureaucracies are likely comprised of relatively more tasks that are
analyzable (less uncertain) and that do not deviate from
757
Bullock
Uncertainty
Low
High
Low
Low Complexity, Low Uncertainty
Low Complexity, High Uncertainty
Task Characteristics:
Few Deviations, High Analyzability
Few Deviations, Low Analyzability
AI Dominates
Leaning AI
Complexity
Relative Strengths of Human and AI
High
Task Characteristics
High Complexity, Low Uncertainty
High Complexity, High Uncertainty
Many Deviations, High Analyzability
Many Deviations, Low Analyzability
Relative Strengths of Human and AI
Leaning Human
Human Dominates
Figure 1. Task complexity, task uncertainty, and discretion.
Note. AI = artificial intelligence.
normal procedures (less complex). Without using Perrow’s
(1967) framework, Bovens and Zouridis (2002) maintain
that factors such as legal and organizational contexts may aid
in the creation of systems-level bureaucracies. Bovens and
Zouridis pose two examples of systems-level bureaucracies
that fit this mold: the monitoring of traffic light violations
and the processing of student loan applications. These two
tasks are more analyzable (less uncertain) and do not deviate
from normal procedures (less complex). It is in this task
environment that systems-level bureaucracies have likely
evolved, and that AI may play the largest role in impacting
human discretion.
AI Influence on Systems-Level
Bureaucracies and Discretion
Systems-level bureaucracies are characterized by ICT tools
that decisively serve the roles of execution, control, and
external communication; typical cases require no human
interference. This means that many tasks have been fully
automated. Systems-level bureaucracies, furthermore, operate within a detailed legal regime where there is no executive
discretion. AI, it should be noted, is an advanced ICT tool
that is efficient, scalable, rapidly diffuses, and increases anonymity and psychological distance from the task (Brundage
et al., 2018). It is high in connectivity and updateability
(Harari, 2018), and its capabilities often exceed those of
human actors. Thus, while the scope of discretion is both curtailing and enabling across bureaucracies (but likely decreasing in overall scope), the legal and organizational contexts in
which systems-level bureaucracies arise are contexts that are
likely dominated by tasks that are analyzable (low in uncertainty) and do not regularly deviate from the normal procedure (low in complexity), which happen to be the same types
of tasks for which AI is likely able to replace human discretion. This suggests that while AI is likely to play a strong role
in limiting human discretion across different types of tasks
and levels of bureaucracies, systems-level bureaucracies
seem poised to first experience both its significant opportunities and its threats to governance and discretion.
Three articles provide important arguments concerning
this relationship. Bovens and Zouridis (2002) list two particular concerns for these new bureaucracies: (a) new discretionary powers that are given to the designers of the systems
that run the systems-level bureaucracy and (b) the potential
rigidity of the systems to individual circumstances. Busch
and Henriksen (2018) assert that the growth in digital discretion seems likely to improve ethical and democratic values
for discretionary administrative tasks but presents challenges
for professional and relational values. Finally, Barth and
Arnold (1999) highlight three potential impacts of AI on
administration, which include potential improvements in
democratic governance, concerns about the role of ethics and
professional competence in discretion, and accountability
across the power balance between legislators and administrators and across control of the AI. These arguments are
explored in further detail below. Given their overlap, Barth
and Arnold (1999) and Busch and Henriksen (2018) are discussed together, with a focus on Barth and Arnold’s (1999)
argument.
According to Bovens and Zouridis (2002), the transition
from street-level bureaucracy to systems-level bureaucracy introduces two concerns for a constitutional state,
new discretionary powers granted to the system designers
and the digital rigidity to individual circumstances. AI is
likely to directly impact both of these. First, AI promises
significant improvements in the digital rigidity of ICT systems. AI systems may be able to both improve upon the
“zenith of rational and legal authority” as discussed by
Bovens & Zouridis and allow for that legal authority to be
more personalized, all the while minimizing typical human
biases against groups of people.3 However, advanced AI
758
systems present an even tougher challenge on the dimensions of supervision and public accessibility. Many AI systems are described as black box processes, or systems
viewed in terms of inputs and outputs without any knowledge of its internal workings, which make supervision and
public accessibility challenging. However, recent work by
AI researchers and legal scholars has begun to provide
some potential remedies to this lack of transparency
(Kleinberg, Ludwig, Mullainathan, & Sunstein, 2019).
These concerns also pose long-term questions about how
best to ensure value alignment (matching AI goals with the
values and goals of its designer) and transparent decisionmaking, a preoccupation of a number of AI scholars,
including Nick Bostrom (2014).
Barth and Arnold (1999) argue that advances in AI will
eventually reach a level at which “machines have the capability to go beyond initial programming and learn to learn”
(p. 334). They state that implicit to this ability is the “ability
to exercise judgment as the situation changes (p. 335).” At
this level of AI, Barth and Arnold believe three themes would
be important to consider:
First, the ability to program machines with values and motives
suggests the potential to improve the rationality of decisions
through tools that can apply a known or specified range of values
or biases (theme of responsiveness). Second, the ability to develop
machines that can sense subtle aspects or changes in the
environment suggests tools that can make political or situational
assessments (theme of judgment). Finally, machines that can learn
to learn independently suggest a tool without precedence that may
exceed the capacity of humans to scan the environment, assess
situations, and make decisions in a timely manner without human
supervision (theme of accountability). (Barth & Arnold, 1999)
Barth and Arnold further argue that AI could improve democratic governance, in that it could be programmed to efficiently
and effectively carry out the values or policies of political
actors. However, the “programmable administrator” (the term
used by Barth & Arnold) eliminates the chance for the personal ethics or professional competence of human bureaucrats
to curb the excesses of these same politicians. On the theme of
judgment, Barth and Arnold reiterate the need to better understand what role professional judgment should play in the
administrative discretion function. Finally, on the theme of
accountability, Barth and Arnold pose two questions:
First, would AI systems lessen the reliance of legislators on the
bureaucracy, or vice versa (and if so, is this desirable?); and
second, does the nature of AI (i.e., the ability to think
independently) suggest systems that are potentially beyond the
control of those responsible for them (i.e., public officials)?
(p. 342)
In response to the first question the authors find that
it does not seem that the increased technical capacity potentially
represented by an AI system would affect the balance of
control between legislators or administrators other than the
American Review of Public Administration 49(7)
potential to reduce the number of analysts on both sides of the
fence. (p. 342)
And as for the second question, along with many leading AI
thinkers, they state that “there is the distinct danger of AI
systems becoming used so much that they become impossible to monitor either because they simply exceed the capacity of humans or the humans lose master of their subject
through lack of use” (p. 342).
Discretion is growing for system designers within systems-level bureaucracies, which can be both humans and AI
designers. This discretion, while dependent upon individual
circumstances, may still lead to more equitable, ethical, and
democratic implementation overall, by allowing for responses
that do not overweight individual circumstances (in the direction of either favoritism or discrimination). AI is likely to
weaken discretion for professional competence and relational
values, which could affect human bureaucrats’ relationships
with citizens and the citizens’ relationships with the state.
Finally, AI systems can be opaque black box processes, with
capacities in some task areas that outpace those of humans,
raising concerns of control and accountability. Given these
concerns, it is important to proceed carefully with the deployment of AI to replace human discretion. As Figure 1 suggests,
for tasks that are lower in both complexity and uncertainty,
these concerns may be minimized and the use of AI is encouraged, but for tasks that are characterized by more complexity
and more uncertainty, the use of AI should be approached
with caution.
AI has been increasing in capacity and thus in overall
intelligence and ability to accomplish discretionary tasks. As
AI capacity exceeds that of humans in many task domains,
this increase in intelligence could increase the overall quality
of discretionary tasks, and thus the overall quality of administration. But even if the overall quality of administration can
be improved, AI changes the nature of risks to good governance in significant and important ways.
Caution and Concern With the Rise of AI
The power of AI should not be underestimated. The Future of
Humanity Institute, OpenAI, and other partners published a
report noting the potential malicious uses of AI (Brundage
et al., 2018). They agree that AI expands existing threats to
governance actors, introduces new threats, and changes the
typical character of threats. Furthermore, they note six security-relevant properties of AI that contribute to its potential
malicious use. AI exhibits a dual-use nature (it has the potential to be used to further both beneficial and nefarious objectives), AI systems are commonly both efficient and scalable,
AI systems can exceed human capabilities, AI systems can
increase anonymity and psychological distance (from the
victim), and AI developments lend themselves to rapid diffusion, and AI systems often contain a number of unresolved
vulnerabilities (Brundage et al., 2018). These properties lead
many to consider AI an existential risk to humanity. An
759
Bullock
existential risk is “one where an adverse outcome would
either annihilate Earth-originating intelligent life or permanently and drastically curtail its potential” (Bostrom, 2002,
p. 2). Organizations such as OpenAI and the Future of
Humanity Institute make it their mission to ensure that AI is
used for the benefit of humanity and not its detriment.
With these risks in mind, it is important to consider which
aims and goals AI should be attempting. Unfortunately, we
lack a complete public ethics framework by which to order
the values that we would direct AI to maximize. Without
such a tool to guide good governance, and given AI’s power
for good and evil, it may be appropriate for public administrators to act in a way that minimizes its potential for administrative evil, or unnecessary harm and suffering inflicted
upon a population (Adams, 2011).
We can appreciate the value of minimizing administrative evil while also realizing that AI will continue to
advance. To stay ahead of the risks and abreast of the
opportunities, we should take stock, as begun here, of
where this technology can viably take us, and its broader
implications for the field of PA. As technology continues to
progress, the number of tasks at the street level that are too
uncertain or too complex for AI will continue to shrink.
This suggests a decreased scope of tasks to be performed by
human bureaucrats, leaving open the question of what this
automation means for organizational outcomes and performance.4 The corollary to this is an overall rise in digital
discretion and digital bureaucracy (Busch & Henriksen,
2018). While it is not the purpose of this article to fully
describe the nature of the transition to digital bureaucracies, it seems obvious that a clear assessment of public values is needed to complete this analysis.
PA has often relied on the professional norms of different
groups of its practitioners to address values-centered questions, but both these professionals and their influence are
likely to decline in the face of rising digital discretion and
digital bureaucracy (Busch & Henriksen, 2018). So how do
we prioritize these values, across multiple types of service
delivery? Should people receive equal treatment under the
law? Or should circumstances be taken into account? What
about the relative importance of effectiveness, efficiency,
and equity? Numerous questions remain about how to maximize the common good while mitigating the possibility of
administrative evil, particularly in the presence of such a
powerful tool of governance.
As more and more tasks continue to be automated within
systems-level bureaucracies, what is the overall impact on
both discretion and performance? The literature on digital
discretion is split, and while evidence suggests there is likely
to be a narrowing of discretion, more theoretical and empirical research is needed to better understand the full effects of
this constriction. In particular, scholars must better understand the competing roles between human systems designers
and their AI counterparts in developing bureaucratic systems
that are effective, efficient, and equitable. One theoretical
suggestion from this essay is that tasks that are low in complexity and uncertainty are the likeliest candidates for automation by AI, whereas tasks high in complexity and
uncertainty should remain under the purview of human discretion. AI should be employed carefully, as it potentially
presents an existential risk to humanity.
This article also suggests that scholars should consider the
effect of AI on the structure of bureaucracy. Bovens and
Zouridis (2002) tracked the shift from street-level to screenlevel to systems-level for certain executive decision-making
organizations. AI is likely to positively impact the performance of organizations that have already become systemslevel bureaucracies, although the extent to which this is true
should be further examined empirically. However, the impact
on bureaucracies that do not fit this characterization is less
clear. The size, resources, and task distribution of these public organizations all affect the likelihood of AI’s influence,
but much more research is needed in this area as well.
Finally, the power and versatility of AI beg the question:
What goals should it be directed to accomplish? In this way,
AI has been described as a dual-use technology: It can be
used toward a wide variety of ends, both good and evil.
Given this, we need a better understanding of which values
of governance should be maximized and the optimal balance
among traditional priorities such as effectiveness, efficiency,
and equity.
The changing character of discretion prompted by the rise
of AI is likely to have important ramifications throughout
society and the economy, but it is just as likely to vitally influence administration and governance. While this article identifies and addresses many of the theoretical and empirical
questions of importance, much more research is needed to
understand the influence of AI on discretion and bureaucratic
performance and to plan a cohesive and effective response.
Conclusion
Justin B. Bullock is also affiliated to Taiwan Institute for
Governance and Communication Research.
Given the intimate relationship between intelligence and
administration, it should not be surprising that AI is encroaching upon the human space of administration and raising a lot
of questions, opportunities, and concerns. AI will continue to
progress in its capabilities, and it will be up to public administrators and PA scholars to discern how best to manage the
rising influence of AI and the waning discretion of streetlevel bureaucrats.
Author’s Note
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect
to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
760
American Review of Public Administration 49(7)
Notes
1.
2.
3.
4.
Given the definition of intelligence, it should be clear that it
can be used to distinguish between human intelligence and
artificial intelligence (AI). Within the fields of computer science, AI, and physics, it is commonly asserted that intelligence
is substrate independent. Substrate independence suggests
that intelligence, or the ability to accomplish a complex task,
can be encoded not just in biological format, as with organic
brains, but in nonbiological format, as with computers.
One striking example of this at a systematic level has been
found by Mogen Pedersen, Justin Stritch, and Frederik Thuesen
(2018), where they find “Overall, our studies support the likelihood that ethnic minority clients will be punished more often
for policy infractions than ethnic majority clients—and that
caseworker work experience mitigates part of this bias”
As noted by multiple scholars, this potential is not always
realized in practice. Virginia Eubanks (2018) makes a compelling argument for how automation of some tasks has led to
decisions that reinforce oppression toward the poor. Andrew
Ferguson (2017) has also highlighted a number of challenges
with predictive policing and “black” data, in which he highlights challenges for discrimination in predictive policing.
For example, Wenger and Wilkins (2008) highlight how automation of the application for unemployment insurance led to
improved outcomes for women.
ORCID iD
Justin B. Bullock
https://orcid.org/0000-0001-6263-2343
References
Adams, G.B. (2011). The Problem of Administrative Evil in a
Culture of Technical Rationality. Public Integrity, 13, 275-286.
Barth, T. J., & Arnold, E. (1999). Artificial intelligence and administrative discretion: Implications for public administration. The
American Review of Public Administration, 29, 332-351.
Bazerman, M. (2002). Judgment in managerial decision making
(5th ed.). Hoboken, NJ: John Wiley.
Borins, S. (1998). Innovating with integrity: How local heroes
are transforming American government. Washington, DC:
Georgetown University Press.
Borins, S. (2008). Innovations in government: Research, recognition, and replication. Washington, DC: Brookings Institution.
Bostrom, N. (2002). Existential risks: Analyzing human extinction scenarios and related hazards. Journal of Evolution and
Technology, 9, 1-30.
Bostrom, N. (2014). Superintelligence. New York, NY: Oxford
University Press.
Bovens, M., & Zouridis, S. (2002). From street-level to systemlevel bureaucracies: How information and communication
technology is transforming administrative discretion and constitutional control. Public Administration Review, 62, 174-184.
Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley,
P., Garfinkel, B., . . . Anderson, H. (2018). The malicious use
of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv: 1802.07228.
Buffat, A. (2011). Discretionary Power and Accountability of the
Office Bureaucracy: The Taxators of a Unemployment Fund as
Implementing Actors (Unpublished doctoral thesis). University
of Lausanne, Switzerland.
Buffat, A. (2015). Street-level bureaucracy and e-government.
Public Management Review, 17, 149-161.
Bullock, J. B., Greer, R. A., & O’Toole, L. J. (2019). Managing
risks in public organizations: A conceptual foundation and
research agenda. Perspectives on Public Management and
Governance, 2, 75-87.
Bullock, J. B., Hansen, J. R., & Houston, D. J. (2018). Sector differences in employee’s perceived importance of income and job
security: Can these be found across the contexts of countries,
cultures, and occupations. International Public Management
Journal, 21, 243-271.
Bullock, J. B., Stritch, J. M., & Rainey, H. G. (2015). International
comparisons of public and private employees’ work motives,
attitudes, and perceived rewards. Public Administration
Review, 73, 479-489.
Busch, P. A., & Henriksen, H. Z. (2018). Digital discretion: A systematic literature review of ICT and street-level discretion.
Information Polity, 23, 3-28.
Davenport, C. (2018, June 29). SpaceX is flying an artificially intelligent robot named CIMON to the International Space Station. The
Washington Post. Retrieved from https://www.washingtonpost.
com/news/the-switch/wp/2018/06/29/spacex-is-flying-an-artificially-intelligent-robot-named-cimon-to-the-internationalspace-station/?noredirect=on&utm_term=.66652e44fdf8
Dreary, I. J., Penke, L., & Johnson, W. (2010). The neuroscience of
human intelligence differences. Nature Reviews Neuroscience,
11, 201-211.
Dunn, D. D. (1997). Politics and administration at the top: Lessons
from down under. Pittsburgh, PA: University of Pittsburgh
Press.
Eubanks, V. (2018). Automating inequality: How high-tech tools
profile, police, and punish the poor. New York, NY: St.
Martin’s Press.
Executive Office of the President. (2016). Artificial ­intelligence,
automation, and the economy. Retrieved from https://­
obamawhitehouse.archives.gov/sites/whitehouse.gov/files/
documents/Artificial-Intelligence-Automation-Economy.
PDF
Executive Office of the President. (2019, February 11).
Maintaining American leadership in artificial intelligence
(E.O. 13859). Retrieved from https://www.federalregister.
gov/documents/2019/02/14/2019-02544/maintaining-american-leadership-in-artificial-intelligence
Ferguson, A. (2017). The rise of big data policing: Surveillance,
race, and the future of law enforcement. New York: New York
University Press.
Goldsmith, S., & Eggers, W. D. (2004). Governing by network.
Washington, DC: Brookings Institution.
Greer, R. A., & Bullock, J. B. (2018). Decreasing improper payments in a complex federal program. Public Administration
Review, 78(1), 14-23.
Gulick, L. (1937). Notes on the theory of organization. In L.
Gulick & L. Urwick (Eds.), Papers on the science of administration (pp. 1-49). New York, NY: Institute of Public
Administration.
Harari, Y. N. (2016). Homo deus: A brief history of tomorrow. New
York, NY: Random House.
Harari, Y. N. (2018, October). Why technology favors tyranny.
The Atlantic. Retrieved from https://www.theatlantic.com/
magazine/archive/2018/10/yuval-noah-harari-technology-tyranny/568330/
761
Bullock
Helfstein, S. & Manley, I.P. (2018, February). Alpha Currents
On the Markets. Morgan Stanley Wealth Management,
Global Investment Committee. Retrieved from https://advisor.morganstanley.com/richard.linhart/documents/field/l/
li/linhart-richard-steven/227_AlphaCurrents___Artifical_
Intelligence_and_Automation.pdf
Isett, K., Mergel, I., LeRoux, K., Mischen, P., & Rethemeyer,
K. (2011). Networks in public administration scholarship:
Understanding where we are and where we need to go.
Journal of Public Administration Research and Theory, 21,
i157-i173.
Jorna, F., & Wagenaar, P. (2007). The “iron cage” strengthened?
Discretion and digital discipline. Public Administration, 85,
189-214.
Kahneman, D. (2011). Thinking, fast and slow. New York, NY:
Farrar, Straus and Giroux.
Kahneman, D., & Tversky, A. (1972). Subjective probability: A
judgment of representativeness. Cognitive Psychology, 3,
430-454.
Kahneman, D., & Tversky, A. (1979). Intuitive prediction:
Biases and corrective procedures. Management Science, 12,
313-327.
Kleinberg, J., Ludwig, J., Mullainathan, S., & Sunstein, C. R. (2019).
Discrimination in the age of algorithms. arXiv:1902.03731.
Klitgaard, R. (1988). Controlling corruption. Berkeley: University
of California Press.
Kurzweil, R. (2005). The singularity is near: When humans transcend biology. New York, NY: Penguin Books.
Kurzweil, R. (2013). How to create a mind: The secret of human
thought revealed. New York, NY: Penguin Books.
Linden, R. M. (1990). From vision to reality: Strategies of successful innovators in government. Charlottesville, VA: LEL
Enterprise.
Lipsky, M. (1980). Street-level bureaucracy: Dilemmas of the
individual in public service. New York, NY: Russell Sage
Foundation.
Meier, K. J., & O’Toole, L. J. (2001). Managerial strategies and
behavior in networks: A model with evidence from U.S. public education. Journal of Public Administration Research and
Theory, 11, 271-293.
Meier, K. J., & O’Toole, L. J. (2002). Public management and
organizational performance: The impact of managerial quality. Journal of Public Administration Research and Theory, 21,
629-643.
Nilsson, N. J. (2009). The quest for artificial intelligence: A history of ideas and achievements. New York, NY: Cambridge
University Press.
O’Toole, L. J., & Meier, K. J. (1999). Modeling the impact of public management: Implications of structural context. Journal of
Public Administration Research and Theory, 9, 505-526.
O’Toole, L. J., & Meier, K. J. (2003). Plus ça change: Public management, personnel stability, and organizational performance.
Journal of Public Administration Research and Theory, 13, 43-64.
O’Toole, L. J., & Meier, K. J. (2011). Public management:
Organizations, governance, and performance. Cambridge,
UK: Cambridge University Press.
Pedersen, M. J., Stritch, J. M., & Thuesen, F. (2018). Punishment
on the frontlines of public service delivery: Client ethnicity and
caseworker sanctioning decisions in a Scandinavian welfare
state. Journal of Public Administration Research and Theory,
28, 339-354.
Perrow, C. (1967). A framework for the comparative analysis of
organizations. American Sociological Review, 32, 194-208.
Perry, J. L. (1996). Measuring public service motivation: An assessment of construct reliability and validity. Journal of Public
Administration Research and Theory, 6, 5-24.
Perry, J. L., & Wise, L. R. (1990). The motivational bases of public
service. Public Administration Review, 50, 367-373.
Piaget, J. (2005). The psychology of intelligence. London, England:
Routledge.
Rainey, H. G. (2014). Understanding and managing public organizations. San Francisco, CA: Jossey-Bass.
Russell, S. J., & Norvig, P. (2016). Artificial intelligence: A modern
approach. Kuala Lumpur, Malaysia: Pearson.
Salamon, L. M. (2002). The tools of Government: A guide to the
New Governance. New York: Oxford University Press.
Simon, H. A. (1947). Administrative behavior: A study of decisionmaking processes in administrative organizations. Oxford,
UK: Palgrave Macmillan.
Simon, H. A. (1997). Administrative behavior: A study of decisionmaking processes in administrative organizations (4th ed.).
New York, NY: Simon & Schuster.
Snellen, I. (1998). Street level bureaucracy in an information age.
In I. Snellen & W. van de Donk (Eds.), Public administration
in an information age: A handbook (pp. 497-508). Amsterdam,
The Netherlands: IOS Press.
Snellen, I. (2002). Electronic governance: Implications for citizens, politicians and public servants. International Review of
Administrative Sciences, 68, 183-198.
Tegmark, M. (2017). Life 3.0: Being human in the age of artificial
intelligence. New York, NY: Alfred A. Knopf.
Topol, E. (2019). Deep medicine: How artificial intelligence can
make healthcare human again. New York, NY: Basic Books.
Vitalis, A., & Duhaut, N. (2004). NTIC and administrative relationship: from the window relationship to the network relationship.
French Review of Public Administration, 110, 315-326.
Weller, J.-M. (2006). Poularde farmer must be saved from remote
sensing: public concern over administrative work. Policies and
Public Management, 24, 109-122.
Wenger, J. B., & Wilkins, V. M. (2008). At the discretion of rogue
agents: How automation improves women’s outcomes in
unemployment insurance. Journal of Public Administration
Research and Theory, 19, 313-333.
Wilkinson, N. (2008). An introduction to behavioral economics.
New York, NY: Palgrave Macmillan.
Zuurmond, A. (1998). From bureaucracy to infocracy: Are democratic institutions lagging behind? In I. T. M. W. Snellen
& B. H. J. van de Donk (Eds.), Public administration in an
information age: A handbook (pp. 259-270). Amsterdam, The
Netherlands: IOS Press.
Author Biography
Justin B. Bullock is an assistant professor of Public Service &
Administration at The Bush School of Government & Public
Service, Texas A&M University. His research interests include
public administration, governance, artificial intelligence, decision
making, discretion, and motivation.